Am I missing something or is it a bit strange to omit the ground truth photos in the visual comparison here?
Off topic, but I love the concept of knowledge distillation.
Cool, but it seems to me like this has limited use. Perhaps some of these ideas could be useful for what I think would be the real prize: a technique to "de-light" a radiance field, recovering/extrapolating the BDRF of each surface so that the scene could be re-lit using path tracing.